Tri Kurniawan Wijaya

As the Director of an advanced AI Research Lab, I lead a talented team of engineers and researchers focused on driving breakthroughs in cutting-edge technologies that are reshaping industries. Our work spans a broad range of AI applications, from improving operational efficiency to enhancing user experience through AI-driven product innovations.

With a focus on cross-functional collaboration, I work closely with engineering, product, and business teams to ensure that our algorithms and solutions align with broader company objectives, ultimately driving measurable improvements such as increased revenue, Click-Through Rate, and Conversion Rate.

In addition to leading internal innovations, I collaborate with leadership to help shape AI strategies that align with long-term business goals. I enjoy guiding teams on how to leverage AI for product innovation and scaling operations, ensuring that technology aligns with growth objectives. Our AI-driven solutions have been deployed across a variety of applications, delivering efficiency and scalable, impactful results.

Alongside my work leading internal AI innovations, I have shared insights with various teams and startups, helping them align AI initiatives with business goals and scale effectively. I enjoy exploring opportunities to contribute to initiatives and organizations looking to harness the transformative power of AI to drive growth and efficiency.

Let’s connect and explore how AI can unlock new opportunities!

Recommender Systems

1. Rs4rs: Semantically Find Recent Publications from Top Recommendation System-Related Venues.
   [Paper] [Website] [Cite]
   Rs4rs is a web application that uses semantic search to help researchers efficiently find recent, relevant papers from top Recommender Systems conferences and journals, improving precision and accessibility compared to traditional scholarly search tools.

2. Exploiting Graph Structured Cross-Domain Representation for Multi-domain Recommendation.
   [Paper] [Code + Dataset] [Cite]
   We present MAGRec, a graph neural network-based method that improves multi-domain recommender systems by learning sequential user interactions across domains, using both domain-guided and general representations to enhance performance and mitigate negative knowledge transfer.

3. Online Meta-Learning for Model Update Aggregation in Federated Learning for Click-Through Rate Prediction.
   [Paper] [Slides] [Video] [Cite]
   In Federated Learning for click-through rate prediction, our online meta-learning method improves performance by adaptively weighting client updates and adjusting learning rates, addressing issues with client heterogeneity and tuning inefficiencies, and significantly outperforms existing methods in convergence speed and final results.

4. RBoard: A Unified Platform for Reproducible and Reusable Recommender System Benchmarks.
   [Paper] [Cite]
   RBoard is a new benchmarking framework designed to improve the reproducibility and reusability of recommender systems research.

5. MM-GEF: Multi-modal representation meets collaborative filtering.
   [Paper] [Cite]
   MM-GEF enhances multi-modal recommendation systems by integrating multi-modal content features with collaborative item-user interactions through early-fusion, leading to improved item representations and better performance compared to existing approaches.

6. FedFNN: Faster Training Convergence Through Update Predictions in Federated Recommender Systems.
   [Paper] [Cite]
   FedFNN accelerates decentralized model training in Federated Learning by predicting updates from unsampled users, achieving up to 5x faster training speeds while maintaining or improving accuracy and consistently outperforming other methods, especially when client availability is limited.

Natural Language Processing

1. Topics as Entity Clusters: Entity-based Topics from Large Language Models and Graph Neural Networks.
   [Paper] [Cite]
   Neural topic modeling using entity-based representations from large language models and graph neural networks, showing improved coherence in extracting thematic structures compared to traditional word-level methods.

2. SPICED: News Similarity Detection Dataset with Multiple Topics and Complexity Levels.
   [Paper] [Dataset] [Cite]
   SPICED is a dataset with seven news topics and four complexity levels, designed to improve news similarity detection models and benchmarks it using various algorithms to address the challenges posed by the diverse nature of news content.

3. STA: Self-controlled Text Augmentation for Improving Text Classifications.
   [Paper] [Cite]
   In NLP tasks with low-data regime, our Self-Controlled Text Augmentation (STA) approach improves training data quality by retaining semantic content and avoiding the pitfalls of both simple heuristics and complex deep learning methods.

Computer Vision

• Quantitative Metrics for Evaluating Explanations of Video DeepFake Detectors.
  [Paper] [Video] [Cite]
  The growing challenge of DeepFake technology, despite advances in detection accuracy, is compounded by the lack of effective explanations for model decisions, which is crucial for proper content moderation; this study introduces new metrics to evaluate and improve the quality and informativeness of these explanations, assessing their impact on both classification and explanation performance.

Network Analytics

1. Cognitive Radio Algorithms Coexisting in a Network: Performance and Parameter Sensitivity.
   [Paper in IEEE] [Cite]
   We evaluate the performance of various cognitive radio algorithms under different decision-making approaches to determine which consistently performs best and how sensitive they are to suboptimal parameter settings, finding that performance varies significantly with the algorithm type and parameter configuration.

2. Mining Complex Activities in the Wild via a Single Smartphone Accelerometer.
   [Paper] [Cite]
   We use energy-efficient accelerometer sensors and hierarchical feature construction to improve the recognition of complex daily activities in real-life settings, comparing early and late fusion mechanisms with promising results.

Schema Matching

1. Minimizing Human Effort in Reconciling Match Networks.
   [Paper] [Cite]
   Schema and ontology matching involves aligning schema attributes with ontology concepts for data integration, but the process is uncertain and often requires costly human intervention; this study proposes a formal model and reconciliation process to reduce such effort by using a network of schemas and Answer Set Programming to optimize necessary user input.

2. SMART: A tool for analyzing and reconciling schema matching networks.
   [Paper] [Cite]
   Schema matching facilitates data integration by aligning attributes across different database schemas, and the Schema Matching Analyzer and Reconciliation Tool (SMART) addresses the challenges of validating and reconciling these matches in schema networks by detecting inconsistencies, enforcing constraints, and guiding experts through the validation and conflict-resolution process.

Smart Grid

Energy Demand Analytics

1. Forecasting Uncertainty in Electricity Demand.
   [Paper] [Poster] [Presentation] [Dataset] [Cite]
   We present GAM², an advanced method for estimating the time-varying conditional variance of Generalized Additive Model (GAM) residuals to better assess and incorporate uncertainty in forecasting electricity demand, as demonstrated through a case study on the French transmission grid operator’s data.

2. Consumer Segmentation and Knowledge Extraction from Smart Meter and Survey Data.
   [Paper + Supp. material] [Cite]
   A framework and index for segmenting smart meter consumption data across various dimensions, enabling applications like tailored tariffs, theft detection, and demand response programs, while tracking behavioral changes and identifying cluster characteristics.

3. Symbolic Representation of Smart Meter Data.
   [Paper] [Cite]
   We address the challenge of managing the large volume and detailed nature of smart meter data by converting it into a symbolic representation, enabling more cost-effective analytics and privacy protection while supporting various machine learning algorithms for classification and forecasting tasks.

4. Cluster-based Aggregate Forecasting for Residential Electricity Demand using Smart Meter Data.
   [Paper] [Supp. material] [Cite]
   We address short-term electricity demand forecasting for residential customers by developing a feature selection process for individual households and using Cluster-based Aggregate Forecasting (CBAF) to improve accuracy, finding that CBAF effectiveness varies with the number of clusters and the size of the customer base.

5. Electricity Load Forecasting for Residential Customers: Exploiting Aggregation and Correlation between Households
   [Paper] [Code] [Cite]
   Recent advancements in smart meters enable real-time analysis of household electricity use, and by applying machine learning techniques to statistical relationships and clustering, we enhance forecasting accuracy and efficiency at both individual and district scales.

6. Online Unsupervised State Recognition in Sensor Data.
   [Paper] [Supp. material] [Cite]
   Dataset: REDD dataset. The original dataset address (redd.csail.mit.edu) is unavailable. Please email me for a personal copy.
   An algorithm that converts sensor data into symbols for efficient anomaly detection, forecasting, and state recognition, reducing data size and resource usage while preserving critical events, and operates online in an unsupervised manner.

7. SmartD: Smart Meter Data Analytics Dashboard.
   [Paper + Supp. material + Review response] [Poster] [Presentation] [Demo videos] [Code] [Dataset] [Cite]
   Smart meters provide real-time energy consumption data for advanced applications like demand response and theft detection, but their large volume and speed present computational challenges, which SmartD addresses by allowing analysts to visualize and estimate typical consumer load profiles based on various factors.

8. A Collaborative Framework for Annotating Energy Datasets.
   [Paper] [Cite]
   A collaborative, web-based framework using gamification to crowdsource appliance usage labeling, aiming to improve energy efficiency by linking human behaviors to energy consumption, while addressing the challenges of obtaining ground truth data for appliance activity in the absence of smart appliances.

9. Estimating Human Interactions with Electrical Appliances for Activity-based Energy Savings Recommendations.
   [Paper] [Poster] [Cite]
   Dataset: previously wiki-energy.org is now Pecan Street. See also this announcement.
   An automated method to detect when household appliances are used based on power consumption, aiming to identify patterns of activity and enable personalized energy-efficient measures in smart homes.

10. Temporal Association Rules For Electrical Activity Detection in Residential Homes.
    [Paper] [Dataset] [Cite]
    A data-driven framework using appliance- and circuit-level power data to mine frequent sequential itemsets, identifying time windows of energy usage patterns to better understand and reduce household electricity consumption through machine learning techniques.

11. Leveraging User Expertise in Collaborative Systems for Annotating Energy Datasets
    [Paper] [Code] [Cite]
    We evaluate expert vs. regular users in annotating energy datasets, offering techniques to identify weak workers and optimize user selection, showing that a small set of carefully chosen tasks can effectively assess expertise and predict accurate crowd-combined annotations.

Demand Response

1. When Bias Matters: An Economic Assessment of Demand Response Baselines for Residential Customers
   [Paper] [Review response] [Code] [Cite]

2. An Economic Analysis of Pervasive, Incentive-Based Demand Response
   [Paper] [Cite]

3. DRSim: A Cyber Physical Simulator for Demand Response Systems
   [Paper] [Cite]

4. Matching Demand with Supply in the Smart Grid using Agent-Based Multiunit Auction
   [Paper + Errata] [Presentation] [Code] [Cite]

5. Effective Consumption Scheduling for Demand-Side Management in the Smart Grid using Non-Uniform Participation Rate
   [Paper + Errata] [Code] [Cite]

6. iDR: Consumer and Grid Friendly Demand Response System
   [Paper] [Cite]

7. Privacy Enhanced Demand Response with Reputation-based Incentive Distribution
   [Paper] [Cite]

8. Methodologies for Effective Demand Response Messaging
   [Paper] [Cite]

9. Crowdsourcing Behavioral Incentives for Pervasive Demand Response
   [Paper] [Cite]
   Complete list of the crowd’s submitted ideas: [link]

PhD Thesis

1. Pervasive Data Analytics for Sustainable Energy Systems
   [Paper] [Cite]

For a complete list of publications, see my Google Scholar.